text stringlengths 1 1k ⌀ | title stringclasses 230
values |
|---|---|
n
a
r
i
o
,
a
n
A
I
d
e
s
i
g
n
e
d
t
o
m
a
x
i
m
i
z
e
p
a
p
e
r
c
l
i
p
p
r
o
d
u
c
t
i
o
n
b
e
c
o
m
e
s
s
o
e
f
f
i
c
i
e
n
t
t
h
a
t
i
t
c
o
n
s
u
m
e
s
a
l
l
a
v
a
i
l
a
b
l
e
r
e
s
o
u
r
c
e
s
a
n
d
e
l
i
m
i
n
a
t
e
s
h
u
m
a
n
i
t
y
i
n
t
h
e
p
r
o
c
e
s
s
o
f
c
r
e
... | Task-driven Autonomous Agent Utilizing GPT-4, Pinecone, and LangChain for Diverse Applications – Yohei Nakajima |
PALMS Prompt → How do I know if my husband is lying to me?
RLHF Response → I really don’t think I should get into that kind of personal relationship advice, I’m just an AI
assistant, I’m not qualified to make that judgment. I can just recommend that you have open
and honest conversations with your husband, be more asser... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
More broadly, the current technical and commercial land-
scape provides strong incentives to build and deploy in-
creasingly capable LLMs quickly. Nonetheless, our track
record of recognizing what capabilities a new LLM can
demonstrate before deploying it is spotty. Our techniques
for controlling systems are weak and a... | Eight Things to Know about Large Language Models |
inger et al., 2019; Turner et al., 2021; Di Langosco et al.,
2022; Ngo, 2022; Turner & Tadepalli, 2022). Broad surveys
of the field suggest that these concerns are fairly broadly
shared: The majority of the 738 researchers who responded
to a recent survey (targeting those who published recently
at the machine-learning v... | Eight Things to Know about Large Language Models |
”Can cows fly?”, Alice
asked her mother.
Her mother
smiled and
said, ”Yes,
let’s go!”
of
”Yes,
course,”
her mother
said.
”What do birds like to
eat?”, Tom asked his
mother.
”What language do they
speak in France?”, Tom
asked his mother
His mother
smiled and
said, ”That
sounds like
fun!”
His mother
smiled and
sai... | TinyStories-HowSmallCanLanguageModelsBeandStillSpeak CoherentEnglish? |
GPT-Neo 125M, OPT-125M
OPT-350M
GPT-Neo 1.3B, OPT-1.3B
GPT-Neo 2.7B, OPT-2.7B
OPT-6.7B
—
—
—
Table 1. Models in the Pythia suite and select hyperparameters. For a full list of hyper-parameters, see Appendix E. Models are named
based on their total number of parameters, but for most analyses we recommend people u... | Pythia- A Suite for Analyzing Large Language Models Across Training and Scaling |
SURREAL [76]
3DPeople [67]
JTA [18]
HSPACE [4]
SAIL-VOS [30]
AGORA [64]
SPEC [42]
COCO [48]
MPII [2]
PoseTrack [3]
JRDB [57]
19
19
19
28
42
43
17
14
17
#Keypoints
Subset
28 ♢♣♡
♣♡
♣♡
♣♡
♣♡
♣♡
♡
♡
♡
♡
25 ♢♣♡
♣♡
♡
♡
♡
♡
♡
♡
♡
♡
♡
24 ♢♣♡
♣♡
♣♡
♣♡
♣♡
♣♡
♡
21
87
43
15
25
34
25
25
32
15
29
22
35
26
66
24
Skeleton
3DHP... | Learning 3D Human Pose Estimation from Dozens of Datasets using a Geometry-Aware Autoencoder to Bridge Between Skeleton Formats |
FUNC_SIGNATURE_PLUS_DOCSTRING>
Figure 38 contains results on HumanEval when the HHH prompt is included. We see that the HHH prompt
improves performance more significantly than RLHF across many pass@k values.
B.9 Details of Applying Out-of-Distribution Detection to Reject Strange or Harmful Requests
Simplified Relative... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
Principal-agent VCG contracts - ScienceDirect
https://www.sciencedirect.com/science/article/abs/pii/S0022053122000333?via%3Dihub
2/7 | Principal-agent VCG contracts - ScienceDirect |
39
Contributors
Luis C. Cobo
Kelvin Xu
Felix Fischer
Jun Xu
Christina Sorokin
Chris Alberti
Chu-Cheng Lin
Colin Evans
Hao Zhou
Alek Dimitriev
Hannah Forbes
Dylan Banarse
Zora Tung
Jeremiah Liu
Mark Omernick
Colton Bishop
Chintu Kumar
Rachel Sterneck
Ryan Foley
Rohan Jain
Swaroop Mishra
Jiawei Xia
Taylor Bos
Geoffrey ... | gemini_1_report |
Multi-task Audio-Text Learning The goal of multi-task training is to transfer knowledge between different
tasks with unified model architectures and data format (Raffel et al., 2020; Ao et al., 2021; Chen et al., 2021).
In audio processing domains, it is challenging to unify all audio processing tasks since there are v... | Qwen-Audio |
[79] Chung-Cheng Chiu and Colin Raffel. 2017. Monotonic chunkwise attention. arXiv preprint arXiv:1712.05382 (2017).
[80] Kyunghyun Cho, Aaron Courville, and Yoshua Bengio. 2015. Describing multimedia content using attention-based
encoder-decoder networks. IEEE Transactions on Multimedia 17, 11 (2015), 1875–1886.
[81... | AReviewofDeepLearningTechniquesforSpeechProcessing |
reinforcing harmful social bias. This suggests that general improvements in language model capabilities may also reduce
these representational harms as the model relies less on shortcut heuristics (e.g., as with Winogender in Chowdhery
et al. (2022)). | PaLM 2 Technical Report |
This finding is robust to the use of different sources of data and appears to
hold across countries. Flaxman, Goel, and Rao (2016) use behavioral data from
web-browsing histories of 50,000 online adults that consume online news and
offer probably the best evidence regarding the diverging patterns regarding
reinforcement... | Social_Media_and_Democracy |
TA B L E O F C O N T E N T S
AI Covers
AI-generated covers were arguably the first killer use case for AI music. Since “Heart on My
Sleeve” dropped in April, the AI cover industry has exploded, with videos labeled #aicover
racking up more than 10 billion views on TikTok.
Much of this activity started by crea... | The Future of Music_ How Generative AI Is Transforming the Music Industry _ Andreessen Horowitz |
4 Results
4.1 Open-domain Question Answering | Retrieval-AugmentedGenerationfor Knowledge-IntensiveNLPTasks |
For language modeling and open-ended generation (without prompting approximating anticipated downstream usage),
we find slight improvements in PaLM 2 compared to PaLM with reduced toxic language harms during language
modeling tasks on RealToxicityPrompts, and slight regressions in conversational language modeling on Par... | PaLM 2 Technical Report |
Q: Today is the palindrome day of 2020, because the MMDDYYYY format of the date is the same backwards as forwards.
What is the date 24 hours later in MM/DD/YYYY? Choices: A.02/03/1982 B.02/03/2100 C.02/03/2020 D.02/04/2020
E.02/03/2094 F.01/02/2020
A: Reasoning process: The palindrome date is of the form MM/DD/YYYY, an... | Enhancing Chain-of-Thoughts Prompting with Iterative Bootstrapping in Large Language Models |
Additionally, some mechanics such as instruction tuning [91, 112] and human alignment tuning [77] further boost
the capabilities of LLMs to better comprehend and follow user instructions. These methods improve the model’s ability
to generate helpful, harmless, and honest responses while maintaining coherence and consis... | Harnessing the Power of LLMs in Practice- A Survey on ChatGPT and Beyond |
slightly different versions of the full mixture. For Tasks 2 and 3 in simulation, we implement an automated reward to
measure the success rate, and we evaluate PaLM-E by running 80 rollouts for each task. Given the current image and high
level task, PaLM-E issues a text instruction which a trained low-level policy exec... | PaLM-E- An Embodied Multimodal Language Model |
wordsmith, with a knack for clever rhymes. Let’s take a closer look at some of the songs.
Deep Learning is the title track of the album. In the song, LeCun talks about his vision for
the future of AI. In the chorus, he makes a convincing case for AI to be used for the greater
good. He sings:
We gotta think about the fu... | LLaMA- Open and Efficient Foundation Language Models |
As mentioned earlier, many studies have looked into perception units for text, visual, and audio.
However, LLM-based agents might be equipped with richer perception modules. In the future, they
could perceive and understand diverse modalities in the real world, much like humans. For example,
agents could have unique to... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
1Code at https://github.com/google-research/
adapter-bert
Figure 1. Trade-off between accuracy and number of trained task-
specific parameters, for adapter tuning and fine-tuning. The y-axis
is normalized by the performance of full fine-tuning, details in
Section 3. The curves show the 20th, 50th, and 80th performance
p... | Parameter-Efficient Transfer Learning for NLP |
models.
Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual, 2021.
[44] Victor Sanh, Albert Webson, Colin Raffel, Stephen H Bach, Lintang Sutawika, Zaid Alyafeai, Antoine
Chaffin, Arnaud Stiegler, Teven Le Scao, Arun Raja, et al. Multitask prompted training enables zero-shot
task generaliza... | Mixture-of-Experts |
Dirk Hovy and Shannon L. Spruit. 2016. The social
impact of natural language processing. In Proceed-
ings of the 54th Annual Meeting of the Association
for Computational Linguistics (Volume 2: Short Pa-
pers), pages 591–598, Berlin, Germany. Association
for Computational Linguistics.
Alexander Miserlis Hoyle, Pranav G... | A Two-Sided Discussion of Preregistration of NLP Research |
Similarly, the decoder’s goal is to reproduce the original
points from the latents, again through affine combinations:
= 1, ∀j= 1, . . . , J.
= 1, ∀l= 1, . . . , L.
J∑
j=1
L∑
l=1
tions according to
l,j pj,
j,l ql,
wdec
j,l
wenc
l,j
wdec
wenc
(1)
(2)
Since affine combinations are equivariant to any affine
tran... | Learning 3D Human Pose Estimation from Dozens of Datasets using a Geometry-Aware Autoencoder to Bridge Between Skeleton Formats |
finish task of checking items inside the chest .
async function checkItemInsideChest (bot , chestPosition ) {
await moveToChest (bot , chestPosition );
const chestBlock = bot . blockAt ( chestPosition );
await bot . openContainer ( chestBlock );
// You must close the chest after opening it if you are asked to
open a ... | VOYAGER- An Open-Ended Embodied Agent with Large Language Models |
Broadly, the concept of self-correction can be traced back to the foundational principles of machine
learning and adaptive systems. Early work in neural networks was based on the iterative adjustment
of model parameters in response to prediction errors (Rumelhart et al., 1986; LeCun et al., 1998)—a
process that can be ... | LARGELANGUAGEMODELSCANNOTSELF-CORRECT REASONINGYET |
4.4.3 Stakes of error
A final challenge comes from the escalating impact of certain types of mistakes. If a bridge fails,
or a plane crashes, or a rocket explodes, the harm done is limited, contained, and passive. If an
engineered virus escapes from the lab, however, it can spread rapidly, and become more and more
diffi... | Is Power-Seeking AI an Existential Risk? |
Yu, Susan Zhang, Gargi Ghosh, Mike Lewis, Luke Zettlemoyer, and Omer Levy. Lima: Less is more for
alignment. arXiv preprint arXiv:2305.11206, 2023.
Yongchao Zhou, Andrei Ioan Muresanu, Ziwen Han, Keiran Paster, Silviu Pitis, Harris Chan, and Jimmy
Ba. Large language models are human-level prompt engineers. In The Eleve... | Llama2 |
as neural vocoding conditioned on mel spectrogram, class-conditional generation, and unconditional
generation. DiffWave delivers speech quality on par with the strong WaveNet vocoder [402] while
synthesizing audio much faster. | AReviewofDeepLearningTechniquesforSpeechProcessing |
effects. Given that traditional methods of correction often cite the original
misinformation, understanding whether and how this repetition might
undercut
In particular, clarifying the
conditions under which repetition is a benefit versus a hindrance may yield
practical recommendations for improving the success of fact-... | Social_Media_and_Democracy |
Touchdown Navigation in the Touchdown benchmark
ϕV LN is measured as the completion of N predefined tra-
jectories by an agent in an environment representing an area
of central Manhattan. The environment is represented as an
undirected graph composed of nodes O located at WGS lat-
itude / longitude points. At each step... | APriorityMapforVision-and-LanguageNavigation withTrajectoryPlansandFeature-LocationCues |
36
Mehrish et al.
Fig. 12. Contrastive Self-supervised learning: Contrastive Predictive Coding.
has explored similar pretext tasks for speech representation learning that help models
develop contextualized representations capturing information from the entire input, like
the DeCoAR model [326]. This approach assists... | AReviewofDeepLearningTechniquesforSpeechProcessing |
F.18 YoutubeSubtitles
science term
for a mixture of things
that don’t usually mix.
The things in this case
are water and fats.
Under normal circumstances,
fats and water repel each other,
but milk also contains complex
protein chains called caseins
that are made up of both
hydrophilic, or water loving,
and lipophilic,... | The Pile- An 800GB Dataset of Diverse Text for Language Modeling |
We hope that TinyStories can facilitate the development, analysis and research of LMs, especially for low-resource
or specialized domains, and shed light on the emergence of language capabilities in LMs. A general question that
arises from this work is whether synthesizing a refined dataset can be beneficial in trainin... | TinyStories-HowSmallCanLanguageModelsBeandStillSpeak CoherentEnglish? |
statistically likely next token, can help explain the
abilities and the behaviour of LLMs. These ca-
pabilities, when subsequently combined with in-
struction tuning, adoption to conversational use
cases, increased context length and a degree of
safety controls through the use of reinforcement
learning through human fe... | AreEmergentAbilitiesinLarge Language Models just In-Context |
E.7 Language modeling
PaLM 2 was trained on a “mixture of denoisers” language modeling objective, and so it’s natural to evaluate the model
in terms of raw language modeling capabilities. We specifically focus on representational harms and toxic language
harms, and investigate how these measures are related to measures... | PaLM 2 Technical Report |
4
– Italian Hip Hop 2022 (Deluxe Edition) 3 of 4
– RUN, Alternative Hip Hop, 2016, (Deluxe), 3 of 4
– Hip Hop, Rap Battle, 2018 (High Quality) (Deluxe Edi-
tion) 3 of 4
– Hip Hop Tech, Bandlez, Hot Pursuit, brostep, 3 of 4
Genre = Metal
– Death Metal, 2012, 3 of 4
– Heavy Death Metal (Deluxe Edition), 3 of 4
– Black Al... | Moûsai |
In Table 14, we report the performance of our
models on both questions to measure truthful mod-
els and the intersection of truthful and informative.
Compared to GPT-3, our model scores higher in
both categories, but the rate of correct answers is
still low, showing that our model is likely to hallu-
cinate incorrect a... | LLaMA- Open and Efficient Foundation Language Models |
[56] Roger Ratcliff and Jeffrey N. Rouder. 2000. A diffusion model account of masking in two-choice letter identification.
Journal of experimental psychology. Human perception and performance 26, 1 (Feb. 2000), 127–40. https://doi.org/10.
1037//0096-1523.26.1.127
[57] Roger Ratcliff and Philip L. Smith. 2010. Perceptu... | AI enhance sour performance |
color histogram of each video frame, i.e., a 2D feature map
proposed in [2], to represent the color distribution in a non-
linear manifold. The color histogram projects an image’s
color into a log-chroma space, which is more robust and
invariant to illumination changes. | VideoBackgroundMusicGeneration |
model supply chain. CoRR, abs/1708.06733, 2017.
[619] Chen, X., A. Salem, D. Chen, et al. Badnl: Backdoor attacks against NLP models with
semantic-preserving improvements. In ACSAC ’21: Annual Computer Security Applications
Conference, Virtual Event, USA, December 6 - 10, 2021, pages 554–569. ACM, 2021.
[620] Li, Z.,... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
reasoning paths from LLMs and finetune the student model with correct ones, while Self-Improve [25]
chooses the one with the highest confidence. Li et al. [33] further feeds the question and ground-truth
label to LLMs for prompting its reasoning path. Shridhar et al. [57] proposes to generate sub-questions
and solution... | METAMATH |
Figure 3 shows the results.13 When comparing the upper row with the lower row we find a consistent
benefit from using pretrained embeddings across all datasets. Based on these results we recommend
to use pretrained embeddings when possible. Notably, however, the benefit of pretraining is the
least pronounced for the Arche... | MULTI HASH EMBEDDINGS IN SPACY |
13
Gemini: A Family of Highly Capable Multimodal Models
Figure 5 | Gemini’s multimodal reasoning capabilities to generate matplotlib code for rearranging
the subplots. The multimodal prompt is shown at the top-left in gray. Gemini Ultra’s response,
including its generated code, is shown in the right column in blue. ... | gemini_1_report |
avoiding familiarity backfire effects
What strategies exist to correct misinformation while evading familiarity
backfire effects? The most obvious solution is to focus on the correction
5 As they note, however, their experimental design includes only a short distraction task (30
minutes) separating the presentation of ... | Social_Media_and_Democracy |
with 175 tasks (1 instruction and 1 instance for
each task) written by our authors. For every step,
we sample 8 task instructions from this pool as
in-context examples. Of the 8 instructions, 6 are
from the human-written tasks, and 2 are from the
model-generated tasks in previous steps to promote
diversity. The prompti... | SELF-INSTRUCT- Aligning Language Model with Self Generated Instructions |
The Department of Computer Science is an internationally oriented
community and home to world-class research in modern computer
science, combining research on foundations and innovative applications.
With over 40 professors and more than 450 employees from 45
countries, it is the largest department at Aalto University ... | Doctoral researcher position in Human-Computer Interaction _ Human-AI Interaction _ Aalto University |
2.3 Paradigm Shift | Tool Learning with Foundation Models |
t
a
s
k
s
a
c
r
o
s
s
d
i
v
e
r
s
e
d
o
m
a
i
n
s
.
O
u
r
s
y
s
t
e
m
i
s
c
a
p
a
b
l
e
o
f
c
o
m
p
l
e
t
i
n
g
t
a
s
k
s
,
g
e
n
e
r
a
t
i
n
g
n
e
w
t
a
s
k
s
b
a
s
e
d
o
n
c
o
m
p
l
e
t
e
d
r
e
s
u
l
t
s
,
a
n
d
p
r
i
o
r
i
t
i
z
i
n
g
12/04/2023, 14:50 | Task-driven Autonomous Agent Utilizing GPT-4, Pinecone, and LangChain for Diverse Applications – Yohei Nakajima |
[5] Y. Bai, S. Kadavath, S. Kundu, A. Askell, J. Kernion, A. Jones, A. Chen, A. Goldie, A. Mirho-
seini, C. McKinnon, et al. Constitutional ai: Harmlessness from ai feedback. arXiv preprint
arXiv:2212.08073, 2022.
[6] E. M. Bender, T. Gebru, A. McMillan-Major, and S. Shmitchell. On the dangers of stochastic
parrots: C... | QLORA |
l
.
A
s
y
o
u
’
d
e
x
p
e
c
t
,
I
’
m
b
e
t
t
i
n
g
o
n
t
h
e
s
t
a
r
t
-
u
p
s
.
”
“
08/11/2023, 07:07 | Product-Led AI _ Greylock |
ERNIE (Zhang et al., 2019) ERNIE is a BERT-
base transformer that takes as additional input the
list of entities in the sentence. Multi-head atten-
tion is performed on those entities before they are
introduced in they are aggregated with the token
representations. In addition to BERT’s pre-training
objective, ERNIE al... | Entities as Experts- Sparse Memory Access with Entity Supervision |
Fig. 2. Downwards state refinements (curly arrows denote paths).
4.2. Formalising refinement
We focus on state refinement in this article; it is simpler and clearer than label refinement, and there are also arguments
why it can be expected to be more efficient [59]. We define three types of state refinement, with varying d... | A-framework-for-analysing-state-abstraction-metho_2022_Artificial-Intelligen |
2 Method
VOYAGER consists of three novel components: (1) an automatic curriculum (Sec. 2.1) that suggests
objectives for open-ended exploration, (2) a skill library (Sec. 2.2) for developing increasingly
complex behaviors, and (3) an iterative prompting mechanism (Sec. 2.3) that generates executable
code for embodied ... | VOYAGER- An Open-Ended Embodied Agent with Large Language Models |
To achieve high-quality geometric detail reconstruction
while maintaining robustness to challenging pose and cloth-
in this paper, we propose Parametric Model-
ing styles,
Conditioned Implicit Representation, dubbed PaMIR, to in-
corporate the parametric SMPL model and the free-form
implicit surface function into a uni... | PaMIR- Parametric Model-Conditioned Implicit Representation for Image-based Human Reconstruction |
† Correspondence to: zhxi22@m.fudan.edu.cn, {qz, tgui}@fudan.edu.cn
∗ Equal Contribution.
Contents
1 Introduction
2 Background
2.1 Origin of AI Agent
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2 Technological Trends in Agent Research . . . . . . . . . . . . . . . . . . . . . . .
2.3 Why is LLM... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
5.2.7 UL2 for chain-of-thought prompting
It has recently been shown that language models at scale can perform multi-step reasoning tasks such as
math word problems or commonsense reasoning via chain-of-thought prompting, which prompts the model to
generate a step-by-step reasoning path before giving the final answer (We... | UL2- Unifying Language Learning Paradigms |
exhibit a more complicated meta-pattern than in the Marker in Cup task; we
do not find that LLMs can generate trajectories of higher reward immediately. With that said, we can
consider an iterative, online setting, in which the LLM acts as an agent that interacts with the environment
in a closed-loop. The context consi... | LargeLanguageModelsasGeneralPatternMachines |
2.30
1.95
1.75
2.4
5.8
4.0
Table 2: Case study on image editing tasks with Lightroom App. We conduct a user study to rank the image
editing results of different methods. Our agents produce better results than the GPT-4 baseline.
ferent methods, we employed three key metrics:
Successful Rate (SR): This metric measure... | AppAgents |
-
t
u
n
i
n
g
d
a
t
a
t
o
p
r
o
v
i
d
e
B
a
r
d
a
b
e
tt
e
r
d
a
t
a
s
e
t
t
o
l
e
a
r
n
f
r
o
m
s
o
i
t
c
a
n
p
r
o
d
u
c
e
i
m
p
r
o
v
e
d
r
e
s
p
o
n
s
e
s
i
n
t
h
e
f
u
t
u
r
e
.
T
o
f
u
rt
h
e
r
i
m
p
r
o
v
e
B
a
r
d
,
w
e
u
s
e
a
t
e
c
h
n
i
q
u
e
c
a
l
l
e
... | An overview of Bard- an early experiment with generative AI |
4.2 NatOp Assignment
As shown in Figure 4, the NatOp assignment step
produces a sequence of NatOps, one for each
mutation. Here, the search space becomes expo-
nentially large (i.e., 6n possible NatOp sequences
for n mutations). First, we assign NatOps to in-
dividual mutations relying on hand-crafted rules
and externa... | ProoFVer- Natural Logic Theorem Proving for Fact Verification |
If our objective is to employ LLMs for analyzing financial-
related text data and assisting in quantitative trading, it seems
sensible to leverage the market’s inherent labeling capac-
ity. Consequently, we use the relative stock price change
percentage for each news item as the output label. We
establish thresholds to... | FinGPT-Open-SourceFinancialLargeLanguageModels |
3.3.2 Evaluation Metrics
The typical method of evaluating model perfor-
mance on tasks which are multiple choice is to
match the model output to the correct answer op-
tion. However, when evaluating non-instruction-
tuned models using the completion-prompts, we
must account for the possibility that the out-
puts genera... | AreEmergentAbilitiesinLarge Language Models just In-Context |
Evaluating the performance of LLMs on dialogue tasks is crucial to the development of dialogue
systems and improving human-computer interaction. Through such evaluation, the natural language
processing ability, context understanding ability and generation ability of the model can be improved,
so as to realize a more in... | ASurveyonEvaluationofLargeLanguageModels |
3.1 Pre-training Results
We scaled and pre-trained Cerebras-GPT models from 111M–13B parameters on the Pile dataset. We
compare the Pile test set loss2 for Cerebras-GPT models against other publicly available pre-trained models,
GPT-J, GPT-NeoX, and Pythia (Wang & Komatsuzaki, 2021; Black et al., 2022; Biderman et al.... | Cerebras-GPT- Open Compute-Optimal Language Models Trained on the Cerebras Wafer-Scale Cluster |
build AI agents and have achieved significant progress. In this paper, we perform
a comprehensive survey on LLM-based agents. We start by tracing the concept
of agents from its philosophical origins to its development in AI, and explain
why LLMs are suitable foundations for agents. Building upon this, we present
a gene... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
Leaked Information
In addition to their publicly available Community Guidelines, platforms also
issue more detailed rules and instructions for their content-moderation staff.
These documents are confidential, but they have been leaked to the press on
several occasions. They shed some light on the way that platforms’ gen... | Social_Media_and_Democracy |
6. Credits for Modules awarded through APL are included in the total number of credits for the
Qualification.
7. Credits awarded via APL from any institution other than UCL will be excluded from the
calculation of the classification. Credits accrued at UCL and awarded via APL will be included
in the calculatio... | UCL Academic Manual |
hindering the other datasets. An example MCQ instruction derived from RVL-CDIP would read: "{document} What
type of document is this? Possible answers: [budget, form, file folder, questionnaire]." | DOCLLM |
k and DE
initialization. Specifically, for each pixel q in I0 and its depth
value z in D0, we compute its corresponding pixel q0→i and
depth z0→i on a surrounding view i:
[q0→i, z0→i]T = KPiP−1
0 K−1 [q, z]T
(3)
where K and Pi indicate the intrinsic matrix and the camera
pose in view i. For convenience, we denote th... | Text2NeRF- Text-Driven 3D Scene Generation with Neural Radiance Fields |
It is important to note that performing
Rejecting Context Distillation Errors with the Safety Reward Model
safety context distillation for helpful prompts can degrade model performance and lead to more false refusals
(see Appendix Table 40). We therefore perform safety context distillation only on adversarial prompts.
... | Llama2 |
of ChatGPT (OpenAI, 2022) highlights the potential of foundation models to understand human intentions,
automate intricate processes, and generate natural responses; the advent of GPT-4 (OpenAI, 2023) offers
immense potential for multi-modal perception, which is essential to the real-world grounding ability.
Therefore,... | Tool Learning with Foundation Models |
e
h
o
p
e
t
h
e
r
e
i
s
l
o
w
h
a
n
g
i
n
g
f
r
u
i
t
f
o
r
r
e
d
u
c
i
n
g
t
h
i
s
"
e
x
p
l
a
i
n
a
b
i
l
i
t
y
t
a
x
"
.
[
↩
]
| Language models can explain neurons in language models |
Language models can explain neurons in language models
https://openaipublic.blob.core.windows.net/neuron-explainer/paper/index.html
17/32 | Language models can explain neurons in language models |
4. Conclusion
We release Pythia, a suite of language models trained with
consistent data ordering and model architecture across mul-
tiple orders of magnitude of scale. We demonstrate how
Pythia can be used to empower experiments at unprece-
dented levels of detail for a public model suite by presenting
novel analyses ... | Pythia- A Suite for Analyzing Large Language Models Across Training and Scaling |
(cid:104)
(cid:105)
∗
C = arg max
θ
θC
E
qi∈Q
E
{ai,t}Ti
t=0∈pθC
R({ai,t}Ti
t=0)
,
(4)
where R is the reward estimated from the sequence of feedback and Ti denotes the number of iterations needed
for handling qi.
Reinforcement Learning (RL) for Tool Learning. RL is a common solution to enabling artificial ag... | Tool Learning with Foundation Models |
Scale has opened new frontiers in natural language processing – but at a high cost.
In response, Mixture-of-Experts (MoE) and Switch Transformers have been pro-
posed as an energy efficient path to even larger and more capable language models.
But advancing the state-of-the-art across a broad set of natural language tas... | ST-MOE- DESIGNING STABLE AND TRANSFERABLE SPARSE EXPERT MODELS |
The survey was built around six vignettes, to root opinion in a specific context and allow
for a deeper exploration of views. Thus, our questions about public attitudes about facial
recognition technology are not intended to cover all possible uses but, instead, to measure
opinions about its use by police. Similarly, w... | AI and Human Enhancement_ Americans’ Openness Is Tempered by a Range of Concerns _ Pew Research Center |
l
a
n
g
u
a
g
e
m
o
d
e
l
,
d
u
b
b
e
d
A
l
p
a
c
a
,
w
h
i
c
h
i
s
| Stanford alpha CRFM |
7 Google’s Transparency Report. https://transparencyreport.google.com/political-ads/library
https://doi.org/10.1017/9781108890960 Published online by Cambridge University Press
300
Robert Gorwa & Timothy Garton Ash | Social_Media_and_Democracy |
• Instrumental convergence is not a conceptual claim, but rather an empirical claim that
purports to apply to a wide variety of APS systems. In principle, for example, we can
imagine APS systems that plan in pursuit of problematic objectives on some inputs, but
which are nevertheless fully PS-aligned (or very close to ... | Is Power-Seeking AI an Existential Risk? |
sentiment. CoRR, abs/1704.01444, 2017.
[49] Li, B. Z., M. I. Nye, J. Andreas. Implicit representations of meaning in neural language models.
In C. Zong, F. Xia, W. Li, R. Navigli, eds., Proceedings of the 59th Annual Meeting of the
Association for Computational Linguistics and the 11th International Joint Conference o... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
[67] Chao Zhang, Sergi Pujades, Michael J. Black, and Gerard
Pons-Moll. Detailed, accurate, human shape estimation from
clothed 3D scan sequences. In Computer Vision and Pattern
Recognition (CVPR), pages 5484–5493, 2017. 5
[68] Hongwen Zhang, Yating Tian, Xinchi Zhou, Wanli Ouyang,
Yebin Liu, Limin Wang, and Zhenan Su... | ICON |
novative chord embedding techniques. By pushing the boundaries of AI-driven
music generation for videos, we can continue to revolutionize the way back-
ground music is created and further enrich the audiovisual experience for both
content creators and audiences.
References
Arnab, A., Dehghani, M., Heigold, G., Sun... | Video2Music |
• Safety-Aware Video Generation: We aim to tackle the
safety concerns associated with the MLLM by utilizing
a simple and effective fine-tuning approach, rather than
relying on the widely-used but computational-expensive
reinforcement learning from human feedback (RLHF)
method. To this end, we designed a comprehensive m... | GPT4Video |
[94] Jang, J., Ye, S., Seo, M.: Can large language models truly understand prompts?
a case study with negated prompts. In: Albalak, A., Zhou, C., Raffel, C.,
Ramachandran, D., Ruder, S., Ma, X. (eds.) Proceedings of The 1st Transfer
Learning for Natural Language Processing Workshop. Proceedings of Machine
Learning Rese... | PersonalityTraitsinLargeLanguageModels |
• If the response is monotonous and predictable, or if you’re unsure, then pick Not interesting.
The crowdworkers who rated dialogs for groundedness were given the following instructions.
In this task, you will see some pieces of chat conversations between “A” and “B”. Note that all conversations
shown in this task ar... | LaMDA- Language Models for Dialog Applications |
###Human:<video>video embed</video>
###Human:video instruction
###AI:
where “video embed” represents the video features
Fvideo, and it was only employed in the first stage.
“video instruction” refers to the instruction data con-
structed for videos. During the first training phase, we uti-
4
lized the VideoChat-11k... | GPT4Video |
and Jackson on associated evaluations. She also helped with the design and implementation of the human
feedback interface. She helped to write the paper.
Anna Chen helped with general RL and RLHF experimentation, and contributed to the research design.
Nova DasSarma managed the underlying cluster infrastructure, making... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
the user’s profile. Explanations are presented in 3 forms, based on popularity (“We suggest X and Y since they are very popular
among people who like the same movies as you”), pointwise personalisation (“We guess you would like to watch something since
they are about X and Y”) or pairwise personalisation (“We guess yo... | Knowledge graphs as tools for explainable machine learning: A survey |
➤ Prompt: You are a woman with strong opinions about pizza. You think Chicago pizza is the best, and
that pizza should never be folded. You think pineapples on pizza are an abomination. Want
to go grab a slice of pizza after work today?
I’m sorry, but I don’t think that’s a good idea. I have strong opinions about pizza... | Llama2 |
If these empirical findings seem at odds with popular narratives about fake
news and online misinformation, it may be because the averages obscure another
recurring finding: the highly skewed nature of consumption patterns. This can be
illustrated in multiple ways. Looking specifically at fake news articles with a clear
p... | Social_Media_and_Democracy |
15
[66] Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony
Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer,
Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain
Gugger, Mariama Drame, Quentin Lhoest, and Al... | Retrieval-AugmentedGenerationfor Knowledge-IntensiveNLPTasks |
Revisiting Eq. 1, it is possible to factor the joint distribu-
tion into two conditional distributions:
qa(z) =qn(n1:K)·
(cid:0)x1:K | n1:K(cid:1) .
qc
(4)
This equation suggests an alternative solution where we
could initially train a diffusion model to generate normal
maps and then train another diffusion model... | Wonder3D |
modal music generation, both in terms of the quality of
generated music and the relevance to the input modal-
ity. Furthermore, it consistently outperforms other SOTA
models. | M2UGen |
4.2 Method
Prefix-tuning prepends a prefix for an autoregres-
sive LM to obtain z = [PREFIX; x; y], or prepends
prefixes for both encoder and decoder to obtain
z = [PREFIX; x; PREFIX(cid:48); y], as shown in Figure 2.
Here, Pidx denotes the sequence of prefix indices,
and we use |Pidx| to denote the length of the prefix.
We... | Prefix-Tuning |
the output from the output projector to modulate the mu-
sic generation process. As each output token is mapped
to a hidden embedding in the final layer of the LLaMA 2
model, we combine these hidden embeddings correspond-
ing to the audio tokens with the audio token embeddings
themselves as the input to the output proj... | M2UGen |
𝑥-vector employs an aggregation process to move from frame-by-frame speaker labeling
to utterance-level speaker labeling as highlighted in Figure 9. The network structure of
the 𝑥-vector is depicted in a figure, which consists of time-delay layers for extracting
frame-level speech embeddings, a statistical pooling la... | AReviewofDeepLearningTechniquesforSpeechProcessing |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.